From Fixed Maintenance Schedules to 15% More Effective Operating Time
A global chemical manufacturer was planning recurring furnace cleaning around fixed run lengths, even though equipment condition varied from unit to unit.
STX Next built a condition-based scheduler using live process data and machine learning forecasts, helping the first site increase effective operating time by approximately 15% while improving maintenance planning across the facility.
A Global Chemical Manufacturer Running Continuous Production
Our client is a global chemical manufacturer operating high-temperature facilities across multiple sites. Its furnaces run continuously, processing feedstock into chemical intermediates used in downstream manufacturing.
During operation, coke deposits gradually build up inside the equipment, reducing heat transfer, raising temperatures, and increasing pressure drop. Because this buildup is unavoidable, the equipment has to be taken offline periodically to remove it.
Premature Cleaning Cost Uptime, While Delays Risked Equipment Damage
The client planned cleaning cycles around fixed run lengths based on historical averages. But individual units could reach critical buildup levels at very different times, even when processing the same feedstock at the same rate. Scheduling the work too early meant giving up usable production time, while waiting too long increased the risk of equipment damage and an unplanned shutdown.
Timing also had to be considered across the entire site. Several production units contribute to the same downstream stage, so taking more than one offline at the wrong time can significantly reduce production. At the same time, the information planners relied on was spread across the process historian, maintenance calendars, static reports, and the experience of senior engineers.
This resulted in:
- Premature cleaning, reducing available operating time
- Delayed cleaning, increasing the risk of equipment damage and unplanned outages
- No single view of unit status, with operators and planners working from different sources
- Manual shift handovers, relying on reports, spreadsheets, and verbal summaries
- No forward-looking estimate, showing when each unit was likely to require service
A Condition-Based Cleaning Scheduler
The team first identified the process signals and thresholds that indicated how deposit buildup was progressing and when each unit was likely to require service.
That information then had to reach planners in a form they could interpret and use when coordinating downtime across the entire site.
STX Next built the scheduler around three connected tracks.
Track 1: Tracking Every Unit From One Live View
The first track brought live process data from the plant historian into one site-wide interface. Operators can see whether each unit is running, being cleaned, on standby, or under maintenance, together with the main signals used to assess its condition:
- Component temperature, a key indicator of deposit buildup
- Pressure drop, which increases as deposits restrict flow
- Feed rate and fuel flow, showing how each unit is operating
- Elapsed run length, compared with the forecast rather than a fixed target
The same view gives incoming and outgoing teams a common reference during shift handovers instead of relying on manually prepared summaries.
Track 2: Forecasting When Each Unit Will Need Cleaning
The forecasting module uses historical process data and current sensor readings to estimate when each unit will reach its cleaning threshold, replacing fixed run-length averages.
The modelling work included:
- Physics-based features developed with the client's process engineer
- A shared architecture designed to work across multiple units instead of maintaining a separate model for each one
- Componentized ML pipelines that made additional unit types easier to add and test
- SHAP explainability showing the factors behind each forecast
Engineers can see what is affecting each prediction alongside its date and confidence range, helping them judge whether the unit needs cleaning or whether an operating issue can be corrected first.
The module also forecasts near-term energy use and expected output, helping planners weigh a longer production run against the additional energy required.
Track 3: Turning Forecasts Into a Site-Wide Maintenance Plan
The scheduler combines predicted service timing for individual units with production targets, existing maintenance windows, and constraints such as crew availability.
It flags units likely to need attention earlier than expected and adjusts the sequence as conditions change. Planners can accept a recommendation, override it with a reason, or leave notes for the next shift. Approved changes are then reflected in the site's execution schedule.
Building the Application Layer
The application also had to handle a new backend provider and continuously updated time-series sensor data. The frontend data layer was refactored around the new source, with frontend and backend changes coordinated as both developed in parallel.
ReactJS was retained to stay consistent with technologies already used by the client. The work also covered UX/UI design and manual and automated testing.
From Reactive Cleaning to More Predictable Site Operations
Roughly 15% More Effective Operating Time
After full rollout at the first site, it recorded approximately 15% more effective operating time through fewer premature cleanings and fewer unplanned outages.
Fewer Reactive Cleaning Events
Earlier forecasts reduced the need to react after problems had already developed and allowed more service work to be coordinated with existing production and maintenance plans.
Lower Cost per Unit of Output
Better timing of service work reduced fuel consumption per ton produced and lowered the risk of costly equipment damage.
Faster Shift Handovers
A shared live view replaced manually compiled status reports, giving operators one screen for reviewing the condition of every unit.
Forecasts With Clear Drivers
Each forecast shows the factors behind it, helping engineers distinguish between a genuine cleaning need and an operating issue that can be corrected.
Site-Wide Maintenance Planning
The scheduler accounts for cleaning forecasts, production targets, crew availability, and existing maintenance windows, reducing the risk of avoidable simultaneous outages.
Turning Process Data Into Operational Decisions
The project brought together information that had previously been spread across separate systems and workflows. Live equipment data, maintenance windows, and production targets now feed into one planning process, while input from the client’s process engineer helped shape the signals, thresholds, and forecasting logic behind it.
Machine learning is only one part of that system. The forecasts become useful because they feed into a scheduler that accounts for site constraints, while the interface shows engineers the factors behind each forecast. Combined with a shared operational view, this turns process data into information operators and planners can use when deciding when equipment should be taken offline and how that downtime should be coordinated across the site.
Ready to make maintenance planning more proactive?
See how data and AI can help you anticipate service needs and plan downtime more effectively.
Our customers love to work with us
